SBIR-STTR Award

This Small Business Innovation Research (SBIR) Phase I project aims at solving the computational problem of personalizing music search and recommendation. The recent explosion of digital music has created an urgent need for powerful knowledge management techniques and tools. Because of the highly subjective nature of musical content and perception, the best possible search strategy would rank media in a personalized fashion, based on each individual's tastes and preferences, from combined cultural and acoustic descriptions. The Echo Nest's predictive personalization technology computes and collects, collaboratively and automatically, cultural opinions online and acoustic content using unsupervised data mining and machine listening techniques. Combining cultural and acoustic notions of music together with the analysis of an individual's listening patterns, ratings and feedback, leads to a vertical search/recommendation engine that knows about content, communities' reaction, and users' preferences. Intelligent music personalization goes beyond search and recommendation. Because the approach is fully autonomous and scalable it can efficiently address the long tail of independent music as well as the Billboard 100; discover artists and niches or predict trends and hits; market indies directly to individuals and optimize aggregators, distributors, and record labels' selection. The Echo Nest engine is the perceptual-media complement to purely text-based search engines and has a significant market potential

This SBIR Phase II project applies data mining and machine learning techniques to both natural language description and Internet link graphs to model communities in order to predict preference, taste and sentiment for different kinds of media (music, TV, online media, video games, books). Current contextual information mining approaches that scan the text on a page for advertisement or recommendation ignore valuable community connections inherent in most self-published Internet discussion. Sentiment and opinion extraction systems operating on full text create challenging language parsing problems are fraught with issues of scale and adaptability. The identification systems can automatically categorize anonymous Internet writers or website visitors into specific demographic communities based on their tastes in many kinds of media. The Phase II research project approaches opinion extraction with a bias-free learning model based on training from known online corpuses that can be adapted to different languages and learns in real time as more data becomes available for high accuracy. Current personalization and marketing approaches either look at the "clickstream" of an anonymous user, leading to equally anonymous recommendations for popular movies and music -- or by scanning a surface-level overview of the text, leading to keyword advertisements with limited contextual understanding of entertainment content and community sentiment. The project plans to fully integrate people-focused community and sentiment analysis technologies into an autonomous, learning and scale-free "media knowledge service" for digital entertainment providers and marketers that can change the way digital content is marketed and sold.